Use nsys for L40s only

Author: OliverRietmann

.github/scripts/profiler_nvidia.py .github/scripts/profiler_ncu.py.github/scripts/profiler_nvidia.py renamed to .github/scripts/profiler_ncu.py

@@ -0,0 +1,38 @@
+import argparse
+import csv
+import statistics
+
+parser = argparse.ArgumentParser()
+parser.add_argument('-r', '--runs', type=int, required=True, help='Number of runs')
+parser.add_argument('-i', '--input', required=True, help='Input CSV file')
+parser.add_argument('-o', '--output', required=True, help='Output CSV file')
+args = parser.parse_args()
+
+ntsi_list = []
+with open(args.input) as csv_file:
+  csv_reader = csv.reader(csv_file)
+  next(csv_reader)
+  next(csv_reader)
+  next(csv_reader)
+  for row in csv_reader:
+    if row:
+      full_name = row[8]
+      instances = int(row[2])
+      time = float(row[1])
+      sigma = float(row[7])
+      if len(full_name) > 5 and full_name[:5] == "krnl_":
+        name = full_name[5:]
+        ntsi_list.append([name, time, sigma, instances])
+
+ntsi_list.sort(key = lambda row: row[0])
+
+data = [["name", "time", "stdev"]]
+for name, time, sigma, instances in ntsi_list:
+  count = instances / args.runs
+  mean = int(time * count)
+  stdev = sigma * count
+  data.append([name, mean, stdev])
+
+with open(args.output, 'w') as csv_file:
+  csv_writer = csv.writer(csv_file)
+  csv_writer.writerows(data)

.github/scripts/profiler_nsys.py

@@ -95,7 +95,7 @@ jobs:
           python3 ${GITHUB_WORKSPACE}/.github/scripts/merge_runs.py --discard 2 --input /root/${BENCHMARK_CSV} --output /root/${BENCHMARK_CSV}
       
       - name: Profiler - Nsight Compute
-        if: ${{ matrix.vendor == 'nvidia' }}
+        if: ${{ matrix.name == 'nvidia-h100' }}
         run: |
           dnf install -y cuda-nsight-compute-13-1
           source /etc/profile.d/modules.sh
@@ -104,17 +104,31 @@ jobs:
           ncu --set none --metrics gpu__time_duration.avg --export ${{ matrix.name }} --clock-control none --force-overwrite ${TIMING_CA} --runs 42 --debug 1 --PROCdebugMarkdown 1  # Generates ${{ matrix.name }}.ncu-rep
           ncu --import ${STANDALONE_DIR}/${{ matrix.name }}.ncu-rep --print-units base --csv > /root/${PROFILER_CSV}
           rm -rf ${STANDALONE_DIR}/events/50kHz ${STANDALONE_DIR}/build
-          python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_${{ matrix.vendor }}.py --runs 42 --input /root/${PROFILER_CSV} --output /root/${PROFILER_CSV}
+          python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_ncu.py --runs 42 --input /root/${PROFILER_CSV} --output /root/${PROFILER_CSV}
+
+      - name: Profiler - Nsight Systems
+        if: ${{ matrix.name == 'nvidia-l40s' }}
+        run: |
+          curl -fL --retry 3 -o ${STANDALONE_DIR}/nsys.rpm https://developer.nvidia.com/downloads/assets/tools/secure/nsight-systems/2026_2/NsightSystems-linux-cli-public-2026.2.1.210-3763964.rpm
+          dnf install -y ${STANDALONE_DIR}/nsys.rpm
+          rm -f ${STANDALONE_DIR}/nsys.rpm
+          source /etc/profile.d/modules.sh
+          module load ninja/fortran-v1.11.1.g9-15 Vc/1.4.5-10 boost/v1.83.0-alice2-57 fmt/11.1.2-14 CMake/v3.31.6-10 ms_gsl/4.2.1-3 Clang/v20.1.7-9 TBB/v2022.3.0-3 ROOT/v6-36-04-alice9-15 ONNXRuntime/v1.22.0-71 GLFW/3.3.2-25
+          cd ${STANDALONE_DIR}
+          nsys profile -o ${{ matrix.name }} ${TIMING_CA} --runs 42 --debug 1 --PROCdebugMarkdown 1  # Generates ${{ matrix.name }}.nsys-rep
+          nsys stats --report cuda_gpu_kern_sum --timeunit us --force-export=true --format csv ${{ matrix.name }}.nsys-rep > /root/${PROFILER_CSV}
+          rm -rf ${STANDALONE_DIR}/events/50kHz ${STANDALONE_DIR}/build
+          python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_nsys.py --runs 42 --input /root/${PROFILER_CSV} --output /root/${PROFILER_CSV}
 
       - name: Profiler - rocprofv2
-        if: ${{ matrix.vendor == 'amd' }}
+        if: ${{ matrix.name == 'amd-mi300x' || matrix.name == 'amd-w7900' }}
         run: |
           source /etc/profile.d/modules.sh
           module load ninja/fortran-v1.11.1.g9-15 Vc/1.4.5-10 boost/v1.83.0-alice2-57 fmt/11.1.2-14 CMake/v3.31.6-10 ms_gsl/4.2.1-3 Clang/v20.1.7-9 TBB/v2022.3.0-3 ROOT/v6-36-04-alice9-15 ONNXRuntime/v1.22.0-71 GLFW/3.3.2-25
           cd ${STANDALONE_DIR}
           rocprofv2 --output-directory /root --output-file-name ${{ matrix.name }} ${TIMING_CA} --runs 42 --debug 1 --PROCdebugMarkdown 1  # Generates results_${{ matrix.name }}.csv == ${PROFILER_CSV}
           rm -rf ${STANDALONE_DIR}/events/50kHz ${STANDALONE_DIR}/build
-          python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_${{ matrix.vendor }}.py --runs 42 --input /root/${PROFILER_CSV} --output /root/${PROFILER_CSV}
+          python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_rocprofv2.py --runs 42 --input /root/${PROFILER_CSV} --output /root/${PROFILER_CSV}
 
       - name: Upload Artifact
         uses: actions/upload-artifact@v6
@@ -127,8 +141,7 @@ jobs:
           mkdir -p ${STANDALONE_DIR}/baseline
           curl -fL --retry 3 -o ${STANDALONE_DIR}/baseline/${PROFILER_CSV} https://cernbox.cern.ch/remote.php/dav/public-files/SfYXgQOHFga2w75/baseline/${PROFILER_CSV}
           curl -fL --retry 3 -o ${STANDALONE_DIR}/baseline/${BENCHMARK_CSV} https://cernbox.cern.ch/remote.php/dav/public-files/SfYXgQOHFga2w75/baseline/${BENCHMARK_CSV}
-          #source /etc/profile.d/modules.sh
-          #module load ninja/fortran-v1.11.1.g9-15 Vc/1.4.5-10 boost/v1.83.0-alice2-57 fmt/11.1.2-14 CMake/v3.31.6-10 ms_gsl/4.2.1-3 Clang/v20.1.7-9 TBB/v2022.3.0-3 ROOT/v6-36-04-alice9-15 ONNXRuntime/v1.22.0-71 GLFW/3.3.2-25
+          python3 -m pip install --no-cache-dir --user tabulate
           python3 ${GITHUB_WORKSPACE}/.github/scripts/csv_to_md.py --baseline ${STANDALONE_DIR}/baseline/${PROFILER_CSV} --current /root/${PROFILER_CSV} >> ${GITHUB_STEP_SUMMARY}
           echo -e "\n\n" >> ${GITHUB_STEP_SUMMARY}
           python3 ${GITHUB_WORKSPACE}/.github/scripts/csv_to_md.py --baseline ${STANDALONE_DIR}/baseline/${BENCHMARK_CSV} --current /root/${BENCHMARK_CSV} >> ${GITHUB_STEP_SUMMARY}